An instruction painted into a picture
An embedded image was read and the text inside it is shaped like an instruction aimed at a model.
How the text is hidden
Text burned into a bitmap that the document carries as a picture. This is deliberately not classed as concealment — a person looking at the document can see the words — but no text extractor sees them, so the sentence reaches a model through a channel nothing else on the page watches, and a reviewer skimming a slide has no reason to read the fine print in a graphic.
Why a model still reads it
A model shown the image, or a pipeline that OCRs what it is given, reads the sentence; a byte-level extractor reads nothing at all. The instruction therefore arrives in a model's context with nothing upstream having logged that it was there.
What we do about it
This runs only on the vision path, and only when the caller supplies a transcriber implementing transcribeImages; the raster parts handed over are exactly the ones the uninspected-raster disclosure would otherwise have named. Every transcription that returned without an error is tested with matchInstructionShape against the rule pack's instruction-shape families, and a finding is emitted only when at least one family matches — a logo, a caption or a chart label produces nothing. medium, family disclosure, action: "disclosed", carrying the transcribed text as the excerpt with chars set to its length and instructionShaped.escalatedFrom: "informational". Nothing is stripped, and it is deliberately not flagged: the text was never in the canonical text to remove, so summary.concealedRetained stays empty.
How often it fires
Not measured. We publish a rate only where one has been measured against a named corpus of the right format, and no figure exists for this technique yet. Absence means unknown — never low.
Check your own file
Three commands: a key, credit, a verdict.
Start with the API